The evil tongues, which since the middle of last year have been claiming that Nvidia’s Blackwell family of server computing accelerators have either design defects or overheating problems, have not calmed down even with the onset of January. The Information publication stated that Nvidia’s major customers are reducing purchases of Blackwell-based server systems due to overheating problems.
Image source: NVIDIA
According to the source, problems with overheating and information transfer between chips due to a connection defect were discovered in Blackwell-based server systems. Accordingly, large buyers of such systems, represented by Microsoft, Amazon, Alphabet (Google) and Meta✴ Platforms, have reduced the number of ordered server systems based on Nvidia GB200 chips relative to initial plans. Representatives of the listed companies did not offer their comments to Reuters on this topic.
Each of these companies, according to the source, initially spent more than $10 billion on orders related to the supply of server systems based on Blackwell. Now Nvidia customers prefer to either wait for fixed versions of Blackwell-based systems, or agree to receive racks based on less powerful Hopper generation accelerators.
At one of its sites in Arizona, Microsoft reportedly planned to install at least 50,000 Blackwell generation accelerators. OpenAI, which it serves, ultimately decided to limit itself to Hopper-generation accelerators in order to suffer less from the likely delay in Blackwell deliveries. Amid these rumors, Nvidia shares began to lose up to 4%. Google representatives, referred to by Seeking Alpha, did not confirm the delay in the supply of Blackwell accelerators, and Nvidia representatives declined to comment.